[llvm] [AMDGPU] Fix s_buffer_load crash for illegal result types (PR #215483)
Arseniy Obolenskiy via llvm-commits
llvm-commits at lists.llvm.org
Sun Aug 16 23:14:18 PDT 2026
================
@@ -0,0 +1,626 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=amdgpu6.00 | FileCheck %s -check-prefixes=GFX67,GFX6
+; RUN: llc < %s -mtriple=amdgpu7.01 | FileCheck %s -check-prefixes=GFX67,GFX7
+; RUN: llc < %s -mtriple=amdgpu8.02 | FileCheck %s -check-prefixes=GFX8
+; RUN: llc < %s -mtriple=amdgpu9.00 | FileCheck %s -check-prefixes=GFX910,GFX9
+; RUN: llc < %s -mtriple=amdgpu10.10 | FileCheck %s -check-prefixes=GFX910,GFX10
+; RUN: llc < %s -mtriple=amdgpu11.00 -amdgpu-enable-vopd=0 | FileCheck %s -check-prefixes=GFX11
+; RUN: llc < %s -mtriple=amdgpu12.00 -amdgpu-enable-vopd=0 | FileCheck %s -check-prefixes=GFX1200
+; RUN: llc < %s -mtriple=amdgpu12.50 -amdgpu-enable-vopd=0 -mattr=-wait-xcnt | FileCheck %s -check-prefixes=GFX1250
+
+; Result types with no direct SBUFFER_LOAD pattern used to ICE in the type
+; legalizer. GlobalISel is not covered: it hits a separate, pre-existing
+; assertion in AMDGPULegalizerInfo::legalizeSBufferLoad for these widths.
+
+define amdgpu_ps void @s_buffer_load_i1(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_i1:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: s_buffer_load_dword s4, s[0:3], 0x0
+; GFX67-NEXT: s_mov_b32 s2, 0
+; GFX67-NEXT: s_mov_b32 s3, 0xf000
+; GFX67-NEXT: s_mov_b32 s0, s2
+; GFX67-NEXT: s_mov_b32 s1, s2
+; GFX67-NEXT: s_waitcnt lgkmcnt(0)
+; GFX67-NEXT: s_and_b32 s4, s4, 1
+; GFX67-NEXT: v_mov_b32_e32 v2, s4
+; GFX67-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_i1:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_and_b32 s0, s0, 1
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_byte v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_i1:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX910-NEXT: s_waitcnt lgkmcnt(0)
+; GFX910-NEXT: s_and_b32 s0, s0, 1
+; GFX910-NEXT: v_mov_b32_e32 v2, s0
+; GFX910-NEXT: global_store_byte v[0:1], v2, off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_i1:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_and_b32 s0, s0, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s0
+; GFX11-NEXT: global_store_b8 v[0:1], v2, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_i1:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: s_and_b32 s0, s0, 1
+; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1200-NEXT: v_mov_b32_e32 v2, s0
+; GFX1200-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_i1:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: s_and_b32 s0, s0, 1
+; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NEXT: v_mov_b32_e32 v2, s0
+; GFX1250-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call i1 @llvm.amdgcn.s.buffer.load.i1(<4 x i32> %desc, i32 0, i32 0)
+ store i1 %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_i4(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_i4:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: s_buffer_load_dword s4, s[0:3], 0x0
+; GFX67-NEXT: s_mov_b32 s2, 0
+; GFX67-NEXT: s_mov_b32 s3, 0xf000
+; GFX67-NEXT: s_mov_b32 s0, s2
+; GFX67-NEXT: s_mov_b32 s1, s2
+; GFX67-NEXT: s_waitcnt lgkmcnt(0)
+; GFX67-NEXT: s_and_b32 s4, s4, 15
+; GFX67-NEXT: v_mov_b32_e32 v2, s4
+; GFX67-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_i4:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_and_b32 s0, s0, 15
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_byte v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_i4:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX910-NEXT: s_waitcnt lgkmcnt(0)
+; GFX910-NEXT: s_and_b32 s0, s0, 15
+; GFX910-NEXT: v_mov_b32_e32 v2, s0
+; GFX910-NEXT: global_store_byte v[0:1], v2, off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_i4:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_and_b32 s0, s0, 15
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s0
+; GFX11-NEXT: global_store_b8 v[0:1], v2, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_i4:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: s_and_b32 s0, s0, 15
+; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1200-NEXT: v_mov_b32_e32 v2, s0
+; GFX1200-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_i4:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: s_and_b32 s0, s0, 15
+; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NEXT: v_mov_b32_e32 v2, s0
+; GFX1250-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call i4 @llvm.amdgcn.s.buffer.load.i4(<4 x i32> %desc, i32 0, i32 0)
+ store i4 %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_v2i1(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_v2i1:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: s_buffer_load_dword s4, s[0:3], 0x0
+; GFX67-NEXT: s_mov_b32 s2, 0
+; GFX67-NEXT: s_mov_b32 s3, 0xf000
+; GFX67-NEXT: s_mov_b32 s0, s2
+; GFX67-NEXT: s_mov_b32 s1, s2
+; GFX67-NEXT: s_waitcnt lgkmcnt(0)
+; GFX67-NEXT: s_and_b32 s4, s4, 3
+; GFX67-NEXT: v_mov_b32_e32 v2, s4
+; GFX67-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_v2i1:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_and_b32 s0, s0, 3
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_byte v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_v2i1:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
+; GFX910-NEXT: s_waitcnt lgkmcnt(0)
+; GFX910-NEXT: s_and_b32 s0, s0, 3
+; GFX910-NEXT: v_mov_b32_e32 v2, s0
+; GFX910-NEXT: global_store_byte v[0:1], v2, off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_v2i1:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_and_b32 s0, s0, 3
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s0
+; GFX11-NEXT: global_store_b8 v[0:1], v2, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_v2i1:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: s_and_b32 s0, s0, 3
+; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1200-NEXT: v_mov_b32_e32 v2, s0
+; GFX1200-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_v2i1:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b32 s0, s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: s_and_b32 s0, s0, 3
+; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NEXT: v_mov_b32_e32 v2, s0
+; GFX1250-NEXT: global_store_b8 v[0:1], v2, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call <2 x i1> @llvm.amdgcn.s.buffer.load.v2i1(<4 x i32> %desc, i32 0, i32 0)
+ store <2 x i1> %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_v3i16(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX6-LABEL: s_buffer_load_v3i16:
+; GFX6: ; %bb.0: ; %main_body
+; GFX6-NEXT: s_buffer_load_dwordx2 s[4:5], s[0:3], 0x0
+; GFX6-NEXT: s_mov_b32 s2, 0
+; GFX6-NEXT: s_mov_b32 s3, 0xf000
+; GFX6-NEXT: s_mov_b32 s0, s2
+; GFX6-NEXT: s_mov_b32 s1, s2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s5
+; GFX6-NEXT: buffer_store_short v2, v[0:1], s[0:3], 0 addr64 offset:4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s4
+; GFX6-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; GFX6-NEXT: s_endpgm
+;
+; GFX7-LABEL: s_buffer_load_v3i16:
+; GFX7: ; %bb.0: ; %main_body
+; GFX7-NEXT: s_buffer_load_dwordx2 s[4:5], s[0:3], 0x0
+; GFX7-NEXT: s_mov_b32 s2, 0
+; GFX7-NEXT: s_mov_b32 s3, 0xf000
+; GFX7-NEXT: s_mov_b32 s0, s2
+; GFX7-NEXT: s_mov_b32 s1, s2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s5
+; GFX7-NEXT: buffer_store_short v2, v[0:1], s[0:3], 0 addr64 offset:4
+; GFX7-NEXT: v_mov_b32_e32 v2, s4
+; GFX7-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; GFX7-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_v3i16:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, s1
+; GFX8-NEXT: flat_store_short v[2:3], v4
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_dword v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX9-LABEL: s_buffer_load_v3i16:
+; GFX9: ; %bb.0: ; %main_body
+; GFX9-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v2, s1
+; GFX9-NEXT: global_store_short v[0:1], v2, off offset:4
+; GFX9-NEXT: v_mov_b32_e32 v2, s0
+; GFX9-NEXT: global_store_dword v[0:1], v2, off
+; GFX9-NEXT: s_endpgm
+;
+; GFX10-LABEL: s_buffer_load_v3i16:
+; GFX10: ; %bb.0: ; %main_body
+; GFX10-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, s1
+; GFX10-NEXT: v_mov_b32_e32 v3, s0
+; GFX10-NEXT: global_store_short v[0:1], v2, off offset:4
+; GFX10-NEXT: global_store_dword v[0:1], v3, off
+; GFX10-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_v3i16:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, s1
+; GFX11-NEXT: v_mov_b32_e32 v3, s0
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX11-NEXT: global_store_b32 v[0:1], v3, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_v3i16:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: v_mov_b32_e32 v2, s1
+; GFX1200-NEXT: v_mov_b32_e32 v3, s0
+; GFX1200-NEXT: s_clause 0x1
+; GFX1200-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX1200-NEXT: global_store_b32 v[0:1], v3, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_v3i16:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mov_b32_e32 v2, s1
+; GFX1250-NEXT: v_mov_b32_e32 v3, s0
+; GFX1250-NEXT: s_clause 0x1
+; GFX1250-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX1250-NEXT: global_store_b32 v[0:1], v3, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call <3 x i16> @llvm.amdgcn.s.buffer.load.v3i16(<4 x i32> %desc, i32 0, i32 0)
+ store <3 x i16> %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_v6i8(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX6-LABEL: s_buffer_load_v6i8:
+; GFX6: ; %bb.0: ; %main_body
+; GFX6-NEXT: s_buffer_load_dwordx2 s[4:5], s[0:3], 0x0
+; GFX6-NEXT: s_mov_b32 s2, 0
+; GFX6-NEXT: s_mov_b32 s3, 0xf000
+; GFX6-NEXT: s_mov_b32 s0, s2
+; GFX6-NEXT: s_mov_b32 s1, s2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s5
+; GFX6-NEXT: buffer_store_short v2, v[0:1], s[0:3], 0 addr64 offset:4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s4
+; GFX6-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; GFX6-NEXT: s_endpgm
+;
+; GFX7-LABEL: s_buffer_load_v6i8:
+; GFX7: ; %bb.0: ; %main_body
+; GFX7-NEXT: s_buffer_load_dwordx2 s[4:5], s[0:3], 0x0
+; GFX7-NEXT: s_mov_b32 s2, 0
+; GFX7-NEXT: s_mov_b32 s3, 0xf000
+; GFX7-NEXT: s_mov_b32 s0, s2
+; GFX7-NEXT: s_mov_b32 s1, s2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s5
+; GFX7-NEXT: buffer_store_short v2, v[0:1], s[0:3], 0 addr64 offset:4
+; GFX7-NEXT: v_mov_b32_e32 v2, s4
+; GFX7-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; GFX7-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_v6i8:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, s1
+; GFX8-NEXT: flat_store_short v[2:3], v4
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_dword v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX9-LABEL: s_buffer_load_v6i8:
+; GFX9: ; %bb.0: ; %main_body
+; GFX9-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v2, s1
+; GFX9-NEXT: global_store_short v[0:1], v2, off offset:4
+; GFX9-NEXT: v_mov_b32_e32 v2, s0
+; GFX9-NEXT: global_store_dword v[0:1], v2, off
+; GFX9-NEXT: s_endpgm
+;
+; GFX10-LABEL: s_buffer_load_v6i8:
+; GFX10: ; %bb.0: ; %main_body
+; GFX10-NEXT: s_buffer_load_dwordx2 s[0:1], s[0:3], 0x0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, s1
+; GFX10-NEXT: v_mov_b32_e32 v3, s0
+; GFX10-NEXT: global_store_short v[0:1], v2, off offset:4
+; GFX10-NEXT: global_store_dword v[0:1], v3, off
+; GFX10-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_v6i8:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, s1
+; GFX11-NEXT: v_mov_b32_e32 v3, s0
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX11-NEXT: global_store_b32 v[0:1], v3, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_v6i8:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: v_mov_b32_e32 v2, s1
+; GFX1200-NEXT: v_mov_b32_e32 v3, s0
+; GFX1200-NEXT: s_clause 0x1
+; GFX1200-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX1200-NEXT: global_store_b32 v[0:1], v3, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_v6i8:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b64 s[0:1], s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mov_b32_e32 v2, s1
+; GFX1250-NEXT: v_mov_b32_e32 v3, s0
+; GFX1250-NEXT: s_clause 0x1
+; GFX1250-NEXT: global_store_b16 v[0:1], v2, off offset:4
+; GFX1250-NEXT: global_store_b32 v[0:1], v3, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call <6 x i8> @llvm.amdgcn.s.buffer.load.v6i8(<4 x i32> %desc, i32 0, i32 0)
+ store <6 x i8> %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_i128(<4 x i32> inreg %desc, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_i128:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: s_buffer_load_dwordx4 s[0:3], s[0:3], 0x0
+; GFX67-NEXT: s_mov_b32 s6, 0
+; GFX67-NEXT: s_mov_b32 s7, 0xf000
+; GFX67-NEXT: s_mov_b32 s4, s6
+; GFX67-NEXT: s_mov_b32 s5, s6
+; GFX67-NEXT: s_waitcnt lgkmcnt(0)
+; GFX67-NEXT: v_mov_b32_e32 v5, s3
+; GFX67-NEXT: v_mov_b32_e32 v4, s2
+; GFX67-NEXT: v_mov_b32_e32 v3, s1
+; GFX67-NEXT: v_mov_b32_e32 v2, s0
+; GFX67-NEXT: buffer_store_dwordx4 v[2:5], v[0:1], s[4:7], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_i128:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: s_buffer_load_dwordx4 s[0:3], s[0:3], 0x0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, s3
+; GFX8-NEXT: v_mov_b32_e32 v4, s2
+; GFX8-NEXT: v_mov_b32_e32 v3, s1
+; GFX8-NEXT: v_mov_b32_e32 v2, s0
+; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_i128:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: s_buffer_load_dwordx4 s[0:3], s[0:3], 0x0
+; GFX910-NEXT: s_waitcnt lgkmcnt(0)
+; GFX910-NEXT: v_mov_b32_e32 v5, s3
+; GFX910-NEXT: v_mov_b32_e32 v4, s2
+; GFX910-NEXT: v_mov_b32_e32 v3, s1
+; GFX910-NEXT: v_mov_b32_e32 v2, s0
+; GFX910-NEXT: global_store_dwordx4 v[0:1], v[2:5], off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_i128:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: s_buffer_load_b128 s[0:3], s[0:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, s3
+; GFX11-NEXT: v_mov_b32_e32 v4, s2
+; GFX11-NEXT: v_mov_b32_e32 v3, s1
+; GFX11-NEXT: v_mov_b32_e32 v2, s0
+; GFX11-NEXT: global_store_b128 v[0:1], v[2:5], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_i128:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: s_buffer_load_b128 s[0:3], s[0:3], 0x0
+; GFX1200-NEXT: s_wait_kmcnt 0x0
+; GFX1200-NEXT: v_mov_b32_e32 v5, s3
+; GFX1200-NEXT: v_mov_b32_e32 v4, s2
+; GFX1200-NEXT: v_mov_b32_e32 v3, s1
+; GFX1200-NEXT: v_mov_b32_e32 v2, s0
+; GFX1200-NEXT: global_store_b128 v[0:1], v[2:5], off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_i128:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: s_buffer_load_b128 s[0:3], s[0:3], 0x0 nv
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mov_b64_e32 v[4:5], s[2:3]
+; GFX1250-NEXT: v_mov_b64_e32 v[2:3], s[0:1]
+; GFX1250-NEXT: global_store_b128 v[0:1], v[2:5], off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call i128 @llvm.amdgcn.s.buffer.load.i128(<4 x i32> %desc, i32 0, i32 0)
+ store i128 %load, ptr addrspace(1) %out
+ ret void
+}
+
+; A divergent offset takes the MUBUF lowering instead of SMEM.
+define amdgpu_ps void @s_buffer_load_i1_divergent(<4 x i32> inreg %desc, i32 %index, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_i1_divergent:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
+; GFX67-NEXT: s_mov_b32 s2, 0
+; GFX67-NEXT: s_mov_b32 s3, 0xf000
+; GFX67-NEXT: s_mov_b32 s0, s2
+; GFX67-NEXT: s_mov_b32 s1, s2
+; GFX67-NEXT: s_waitcnt vmcnt(0)
+; GFX67-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX67-NEXT: buffer_store_byte v0, v[1:2], s[0:3], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_i1_divergent:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX8-NEXT: flat_store_byte v[1:2], v0
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_i1_divergent:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
+; GFX910-NEXT: s_waitcnt vmcnt(0)
+; GFX910-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX910-NEXT: global_store_byte v[1:2], v0, off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_i1_divergent:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX11-NEXT: global_store_b8 v[1:2], v0, off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_i1_divergent:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen
+; GFX1200-NEXT: s_wait_loadcnt 0x0
+; GFX1200-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX1200-NEXT: global_store_b8 v[1:2], v0, off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_i1_divergent:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen nv
+; GFX1250-NEXT: v_mov_b32_e32 v3, v2
+; GFX1250-NEXT: v_mov_b32_e32 v2, v1
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: v_and_b32_e32 v0, 1, v0
+; GFX1250-NEXT: global_store_b8 v[2:3], v0, off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call i1 @llvm.amdgcn.s.buffer.load.i1(<4 x i32> %desc, i32 %index, i32 0)
+ store i1 %load, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_ps void @s_buffer_load_i128_divergent(<4 x i32> inreg %desc, i32 %index, ptr addrspace(1) %out) {
+; GFX67-LABEL: s_buffer_load_i128_divergent:
+; GFX67: ; %bb.0: ; %main_body
+; GFX67-NEXT: buffer_load_dwordx4 v[3:6], v0, s[0:3], 0 offen
+; GFX67-NEXT: s_mov_b32 s2, 0
+; GFX67-NEXT: s_mov_b32 s3, 0xf000
+; GFX67-NEXT: s_mov_b32 s0, s2
+; GFX67-NEXT: s_mov_b32 s1, s2
+; GFX67-NEXT: s_waitcnt vmcnt(0)
+; GFX67-NEXT: buffer_store_dwordx4 v[3:6], v[1:2], s[0:3], 0 addr64
+; GFX67-NEXT: s_endpgm
+;
+; GFX8-LABEL: s_buffer_load_i128_divergent:
+; GFX8: ; %bb.0: ; %main_body
+; GFX8-NEXT: buffer_load_dwordx4 v[3:6], v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_dwordx4 v[1:2], v[3:6]
+; GFX8-NEXT: s_endpgm
+;
+; GFX910-LABEL: s_buffer_load_i128_divergent:
+; GFX910: ; %bb.0: ; %main_body
+; GFX910-NEXT: buffer_load_dwordx4 v[3:6], v0, s[0:3], 0 offen
+; GFX910-NEXT: s_waitcnt vmcnt(0)
+; GFX910-NEXT: global_store_dwordx4 v[1:2], v[3:6], off
+; GFX910-NEXT: s_endpgm
+;
+; GFX11-LABEL: s_buffer_load_i128_divergent:
+; GFX11: ; %bb.0: ; %main_body
+; GFX11-NEXT: buffer_load_b128 v[3:6], v0, s[0:3], 0 offen
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: global_store_b128 v[1:2], v[3:6], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX1200-LABEL: s_buffer_load_i128_divergent:
+; GFX1200: ; %bb.0: ; %main_body
+; GFX1200-NEXT: buffer_load_b128 v[3:6], v0, s[0:3], null offen
+; GFX1200-NEXT: s_wait_loadcnt 0x0
+; GFX1200-NEXT: global_store_b128 v[1:2], v[3:6], off
+; GFX1200-NEXT: s_endpgm
+;
+; GFX1250-LABEL: s_buffer_load_i128_divergent:
+; GFX1250: ; %bb.0: ; %main_body
+; GFX1250-NEXT: global_prefetch_b8 v0, null scope:SCOPE_SE
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: v_mov_b32_e32 v5, v2
+; GFX1250-NEXT: v_mov_b32_e32 v4, v1
+; GFX1250-NEXT: buffer_load_b128 v[0:3], v0, s[0:3], null offen nv
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: global_store_b128 v[4:5], v[0:3], off
+; GFX1250-NEXT: s_endpgm
+main_body:
+ %load = call i128 @llvm.amdgcn.s.buffer.load.i128(<4 x i32> %desc, i32 %index, i32 0)
+ store i128 %load, ptr addrspace(1) %out
+ ret void
+}
+
+declare i1 @llvm.amdgcn.s.buffer.load.i1(<4 x i32>, i32, i32)
----------------
aobolensk wrote:
Removed
https://github.com/llvm/llvm-project/pull/215483
More information about the llvm-commits
mailing list